On the use of artificial neural network to predict denoised speech quality

Existing objective criteria for denoised speech assessment have as output one score indicating the quality of processed speech. Even it is well useful when it is about comparing denoised techniques between each others, they failed to give with enough accuracy an idea about the real corresponding Mean Opinion Score rate (MOS). In this paper, we propose a new methodology to estimate MOS score of denoised speech. Firstly, a statistical study of existed criteria based on boxplot and Principal Component Analysis (PCA) analysis yields to select the most relevant criteria. Then, an Artificial Neural Network (ANN) trained in selected objective criteria applied on the denoised speech is used. Unlike traditional criteria, the proposed method can give a significant objective score directly interpreted as an estimation of real MOS score. Experimental results show that the proposed method leads to more accurate estimation of the MOS score of the denoised speech.